What We Will Cover Today 👀

We will go backward: from the final reports to the construction of a vibe coding environment.

• A complete Vibe Coded report
• Some limitations and pitfalls
• How to run the Vibe Coded
• How to prepare for Vibe Coding
• Introduction to two main Claude Code features: claude.md and skill.md
• Two Claude Code settings: privacy and usage
• Claude Code usage and cybersecurity

The Take-Home Messages 🧠

• Not faster, but a different way of working.
• Knowledge management.
• Think like a manager: decompose work into well-defined departments.
• It’s not pipelining or automation.
• It probably writes nicer code than you do.

Not Faster

It’s Not Pipelining/Automation

We will write many guidelines for our AI agent (Claude Code). At some point, it might resemble a pipeline.

The key difference is that the details you provide about what you are doing can dramatically change the course of the analysis. Also, you are less constrained by the developer’s pipeline and much more flexible, especially regarding the input.

My rule of thumb:

• For well-controlled data flow, with well-defined input and output formats: build a pipeline. (industry)
• If you want to explore your data with many unanticipated iterations and maximum flexibility: use vibe coding. (science)

Vibe Coding is not replacing anything, it is a new way of working.

The Reports 📊

I asked to create Quarto notebooks with Python as the programming language.

If you’re not familiar with Quarto, it is the successor to RMarkdown. In my opinion, it is far superior to Jupyter. Even these slides were created using Quarto.

Want to learn more?
Please visit Mickaël Canouil’s course, “Mastering Quarto CLI: From Authoring to Publishing,” at https://m.canouil.dev/mastering-quarto-cli/

The report.

Home Page (index.qmd)

We obtain a very nice introduction, it rephrased the project, isolated the questions.

Then we have a nice Study Design section that helps to understand the dataset.

However, here it states the study is IRB-approved. Of course it is, but I didn’t tell it in the context and it had no access to the original paper.

Hallucinations can be very tricky to identify!

Did you notice the trick I used in my prompt?

• It mentions that it is a tutorial. Okay, it is… but the analysis actually isn’t, so why?
• Because I asked it to… but why did I?
• Because I wanted a lot of explanations of what it was doing. In my personal experience, telling it that it is teaching material results in quite verbose but very clear reports.

Data Preprocessing

2 Metabolomics Data Preprocessing.

See “Human notes”:

• Bad coding practice
• Bad imputation practice
• PQN vs autoscale

EDA

• The Timepoint PCA with Plotly really helps with interpretation.

Compound Detection

• You should not only review the Quarto file but also the .py file.
  • Claude independently found the m/z values of Diphenhydramine, N-desmethyl-DPH, and DPH N-oxide.
  • It took me only a few minutes to verify the values with Perplexity.

Multivariate

• We can find N-desmethyl-DPH 4312 in the VIP.
• We can find DPH N-oxide 5053 in the VIP.
• We can find Diphenhydramine 4589 in the VIP.

Where Is Claude? 🗺

Speaking of Claude, we must be clear about what we are talking about because there is quite a bunch of interfaces:

• Web User Interface
• Desktop app
• VS Code extension
• Claude CLI

Claude in a browser 🕸

Claude Desktop 💻

Claude VS Code extension ⌨

Claude CLI 📟

Claude CLI

We want maximum control over what we are doing; thus, Claude CLI is the best choice.

However, from time to time, I use the others for convenience. Here, we will use only Claude CLI.

Launch claude with:

claude

Follow Your Usage (Aka: Credits)

You should keep a close eye on your credits either with usage or on the web page https://claude.ai/settings/usage

Web

Plan Mode: Explore Before You Change

The challenge: Sometimes you need to understand the codebase before making changes

The solution: Plan Mode - read-only exploration first

Type /plan or press Shift+Tab twice to enter Plan Mode:

• Read-only operations: No accidental changes during exploration
• Explore codebase: Search for files, read code, understand structure
• Understand datasets: Check formats, column names, existing preprocessing
• Design approach: Plan implementation strategy
• Get approval: Review plan before execution begins

Switch to plan mode, type /plan or press Shift+Tab twice.

 ▐▛███▜▌   Claude Code v2.1.42
▝▜█████▛▘  Sonnet 4.5 · Claude Pro
  ▘▘ ▝▝    ~\chemometrics_vibe_coding

❯ /plan
⎿  Enabled plan mode
───────────────────────────────────────────────────────────────────────
❯
───────────────────────────────────────────────────────────────────────
⏸ plan mode on (shift+tab to cycle)      

Switch to Opus 4.6 with command /model.

 ▐▛███▜▌   Claude Code v2.1.42
▝▜█████▛▘  Sonnet 4.5 · Claude Pro
  ▘▘ ▝▝    ~\chemometrics_vibe_coding

❯ /plan 
  ⎿  Enabled plan mode

❯ /model
───────────────────────────────────────────────────────────
Select model                                                                                                                                                                                     
 Switch between Claude models. Applies to this session and future Claude Code sessions. For other/previous model names, specify with --model.                                                     

   1. Default (recommended)  Opus 4.6 · Most capable for complex work
 ❯ 2. Sonnet ✔               Sonnet 4.5 · Best for everyday tasks
   3. Haiku                  Haiku 4.5 · Fastest for quick answers  

 ▌▌▌ Effort not supported for Sonnet

 Use /fast to turn on Fast mode (Opus 4.6 only). Now 50% off through Feb 16.

 Enter to confirm · Esc to exit

Opus 4.6 burns through your credits like crazy 🔥. For the workshop, I don’t want to run out of credits, so I’m using medium effort.

The effort parameter allows you to control how eager Claude is to spend tokens when responding to requests. This lets you trade off between response thoroughness and token efficiency, all in a single model.

For the implementation, we will switch to Sonnet 4.5 to save some credits.

First Prompt

Ideally, YOU have an opinion on what is missing from the analysis, and we can try a live prompt.

If not, let’s say we have forgotten to implement the last phase of our initial plan.

The prompt:

@data/CD-MRG-metabolomics_summerschool_2023/plan.md suggest 
another analysis I could do with this dataset

Use @ to reference a file you absolutely want it to read.

Use Ctrl+O to see more details about what is going on.

Context exploration uses a lot of credits.

It identifies that Phase 6 was not completed.

The messier your project is, the more useless files it reads and wastes your credits for nothing.

In addition, it increases the chances of hallucinations.

It proposes various options, we will do Phase 6 implementation

Eventually, it consumed about 20% of my 5-hour credits.

Phase 6 Plan

In a previous prompt I simply asked to plan the 6th phase.

I save the plan to be able to resume later.

Let’s analyze this plan together.

Phase 6 plan

Implement the Plan

I switched to Sonnet 4.5 to save credits.

I planned in a previous session, now I want to implement the @data/CD-MRG-metabolomics_summerschool_2023/plan_phase6.md plan

It runs the code by itself and detects an error.

Then it fixes the error by itself.

Finalize the implementation.

It didn’t implement the quarto notebook, we ask it.

You didn’t create the quarto file for this analysis !

Let’s Look at the Report

Report

Now we saw how Claude behaves when we send simple prompts. What did we do in anticipation to reach such a behavior?

Summary

• Always plan before implementing.
• CAREFULLY read the plan, or you will waste credits and time.
• Run the implementation.
• Review .py and .qmd files not for errors but for sabotage.

Overview

• CLAUDE.md lists what it should always know about the project
• Skills guide Claude on how to perform specific tasks, e.g., “how to preprocess LC-MS data?”
• quarto-authoring Full documentation on Quarto, by Mickaël Canouil
• chemometrics-ms-metabolomics based on Boccard & Rudaz 2018 - Extracting Knowledge from MS Clinical Metabolomic Data
• skill creator a skill to create skill, by Anthropic.

What Is CLAUDE.md?

A markdown file in your project root that Claude Code reads automatically.

It provides:

• Project context (what are we analyzing?)
• Data structure (where are files? what format?)
• Requirements (how to split data? what metrics?)
• Conventions (code style, output locations)
• Constraints (what NOT to do)

These instructions, I provided them in the initial prompt because I wanted to have a user experience like:

• drop your files in “./data”
• write a prompt specific to your use-case and run it
• How can we actually reach this?

## Context

Currently, this repository is optimized for vibe coding with some predefined skills to help.
We expect the user to provide you some data in `./data`, tell you the structure, the context of the analysis and the questions it wants to ask.
As soon as you have all these information, start by suggesting:
- a `CLAUDE.md` edition where this paragraph is removed and replaced by:
- The context of the data
- The data structure
- The main questions and objectives of their analysis.
   - This section should be a checklist regularly updated based on analysis you already did `- [x]`  and new scientific questions `- [ ]`.
   - Feel free to challenge and propose rephrasing of scientific questions
   - Feel free to suggest scientific questions based on context and avaialble data.

It bootstraps the project context, then instructs Claude to replace it with real project details.

Adding Data Structure

## Data Structure

- `data/preprocessing_result.pkl`: Preprocessed metabolomics data
  - Keys: `forearm`, `forehead` (sample locations)
  - Each contains: `X_processed`, `metadata`, `feature_metadata`,
    `peak_areas_raw`, `feature_ids`
  - Features: m/z values (ppm tolerance: ±5 ppm)
  - Timepoints: 0, 1, 2, 4, 8, 12, 24 hours post-dose

- `data/GNPS_results/`: Compound annotation databases
  - GNPS library matches with m/z and retention time

- Outputs: Quarto reports in `reports/`, figures in `_book/`

Adding Common Pitfalls

## Common Pitfalls to Avoid

❌ DO NOT use blank subtraction without careful review
❌ DO NOT use autoscaling (use PQN normalization instead)
❌ DO NOT fit preprocessing on full dataset before analysis
❌ DO NOT ignore mass accuracy and retention time variation
❌ DO NOT rely on statistical significance alone (check biological relevance)
❌ DO NOT report false compounds (verify with GNPS MS/MS data)

✅ Always document feature matching tolerance criteria
✅ Always check multivariate model diagnostics
✅ Always generate dummy datasets with known, controlled effects as a safeguard

SKILL.md Structure

---
name: your-skill-name
description: Brief description of what this Skill does and when to use it
---

# Your Skill Name

## Instructions
[Clear, step-by-step guidance for Claude to follow]

## Examples
[Concrete examples of using this Skill]

---
name: skill-name
description: Brief description of what this Skill does and when to use it
---

The key features here are:

• name: Eventually, you will invoke with /skill-name (or be automatically invoke). Make it handy.
  • Antipattern: choose-methodology-for-machine-learning-in-lcms-context
• description: claude knows this description, it uses it to decide to invoke the skill or not. “should include both what the Skill does and when Claude should use it.”

Only add context Claude doesn’t already have. Challenge each piece of information:

• “Does Claude really need this explanation?”
• “Can I assume Claude knows this?”
• “Does this paragraph justify its token cost?”

In other words: introduce skills either when you want to enforce a behaviour or that Claude failed you previously.

Good example: Concise (approximately 50 tokens):

## Extract PDF text 

Use pdfplumber for text extraction:

```python
import pdfplumber 
   with pdfplumber.open("file.pdf") as pdf: 
      text = pdf.pages[0].extract_text()
```

Bad example: Too verbose (approximately 150 tokens):

## Extract PDF text 

PDF (Portable Document Format) files are a common file format that contains text, 
images, and other content. To extract text from a PDF, you'll need to use a library. 
There are many libraries available for PDF processing, but we recommend pdfplumber 
because it's easy to use and handles most cases well. First, you'll need to install 
it using pip. Then you can use the code below...

Control Autonomy (Degree of Freedom)

---
name: chemometrics-ml-selection
description: Choose the proper normalization method for LC-MS.
---
# Normalization method for LCMS
## High level of freedom
Choose one of: TSN, MSTUS, PQN, LOESS, QC-RSC, QC-SVRC
## Medium freedom
Prefer PQN in most situations.
## Low freedom
Use PQN function to normalize
``` python
def pqn_normalize(X: np.ndarray, reference: np.ndarray | None = None) -> np.ndarray:
    quotients = X / reference
    norm_factors = np.nanmedian(quotients, axis=1, keepdims=True)
    return X / norm_factors
```

Read the End of Doc First

The Claude Skill documentation ends with a checklist.

You can consider it the short version of the documentation, but in any case, it is always better to read it thoroughly.

How Claude Code Uses Skills

When you invoke a skill:

1. Claude Code reads the SKILL.md file
2. Incorporates guidance into its context
3. Applies best practices automatically

Force loading of the skill

# In Claude Code
> /chemometrics-ml-selection

Claude will invoke the skill if it considers it useful in the context

# In Claude Code
> Carefully choose the best ML method

The Skill to Create Skills

The community highly recommend to use /skill-creator when you try to build your own skill.

Build Skill on Top of Scientific Papers

Find a methodology paper you like and agree with like Boccard & Rudaz 2018 - Extracting Knowledge from MS Clinical Metabolomic Data: Processing and Analysis Strategies.

In claude:

Here is a very interesting paper @assets/boccard2018.pdf with a full methodology 
on Metabolomic data analysis. Build a skill with /skill-creator based on this
scientific paper.

Result here: Skill based on Boccard & Rudaz, 2018

Nice, isn’t it?

Can you identify any caveats?

The worst

• We don’t care about references in the skill, it’s a waste of token. Initial prompt should be improved.

• Very low degree of freedom for the multivariate analysis. In the whole skill in general. I should improve my prompt.

• Boccard & Rudaz are very vague about missing values, still the skill has a list of method, not what I would recommend.

I would recommend to read carefully then perform a sanity check with a prompt like

A new paper @path/to/paper just came out on metabolomics statistical analysis. 
Please confront it to my skill /chemometrics-ms-metabolomics.
Create a quarto notebook with the differences you spot.

Build Skill on Top of Scientific Papers

Now that we’ve seen some caveats, what would you do?

Do ✅

• Start with the official skills from Anthropic - they’re battle-tested and well-documented
• Keep your skill collection lean - only enable what you actually need for your current work
• Write clear SKILL.md files with concrete examples showing Claude exactly how to use your skills
• Test new skills on small tasks first before relying on them for complex workflows
• Review skill usage in Claude Code’s output to verify they’re being applied correctly

Don’t 💣

• Download skills from untrusted sources without reviewing their contents first
• Load too many skills at once - each one consumes tokens and slows down execution
• Skip reading the SKILL.md documentation - Claude needs to view it to use skills properly
• Create overly broad skills that try to handle multiple unrelated tasks - keep them focused
• Assume skills are “free” - they consume context on every task, even when not actively used

Skills vs CLAUDE.md

Your Journey Forward

• Write CLAUDE.md for your next project
• Reuse skills on the web, identify bottleneck
• Create your own skills
• Identify your workflow, implement agents
• Share with your team/community
• Iterate and improve

Thank You!

Contact:

• GitHub: AlbanOtt
• LinkedIn: Alban Ott

Workshop materials:

https://github.com/AlbanOtt/chemometrics_vibe_coding_public